In the traditional pipeline of AI-driven drug discovery, a generative model "dreams up" a molecular graph with optimized properties, which is then handed to a chemist to synthesize. This often leads to a frustrating bottleneck: many of these "miracle molecules" are chemical hallucinations—structurally valid on paper but impossible to actually build in a laboratory.
To solve this, a new paradigm has emerged. Instead of generating molecules atom-by-atom, researchers are training models to generate synthesis recipes. By restricting the AI’s action space to valid chemical reactions and commercially available building blocks, these models ensure that every discovered molecule is synthesizable by construction.
From Atom Space to Reaction Space
Traditional generative models operate in "atom space," adding one atom or bond at a time to form a graph. While flexible, this approach has no inherent understanding of chemical feasibility. In contrast, models like RGFN (Reaction-GFlowNet), developed by Michał Koziarski and Yoshua Bengio, shift the generative process into "reaction space."
In this framework, the model starts with a library of purchasable building blocks and applies a sequence of valid reaction templates (such as amidation or Suzuki coupling). As Miruna Cretu and Charles Harris demonstrate with SynFlowNet, this ensures that the final product is not just a theoretical graph, but the end result of a documented chemical pathway. By "thinking" in terms of reactions, the model inherently respects the laws of the wet lab.
The Synthesis Tree as a Data Structure
To navigate this space, researchers represent molecules as synthesis trees. In these trees, the leaf nodes are starting materials and the internal nodes are chemical reactions.
Alston Lo and Connor W. Coley utilize this in SynGA, a genetic algorithm that evolves these "recipes" rather than the molecules themselves. Similarly, Aryan Pedawi and colleagues at Atomwise developed CSLVAE, which acts as a "neural database" for trillion-scale libraries. By encoding the hierarchical rules of synthesis directly into the VAE architecture, they can navigate massive combinatorial spaces with 100% synthetic accessibility.
This approach transforms the search problem. Instead of exploring the nearly infinite space of all possible graphs, the model explores the much smaller (though still vast) space of makeable graphs.
Projecting Dreams into Reality
Sometimes, we already have a "dream molecule" from an unconstrained model or a protein-docking simulation. The challenge then is to find the closest version of that molecule that can actually be manufactured.
This is known as synthesizable projection. Shitong Luo and Jian Peng proposed a framework in Projecting Molecules into Synthesizable Chemical Spaces that uses a transformer to map abstract graphs onto concrete, step-by-step recipes. Similarly, MoleBridge, developed by Rongchao Zhang and others, uses a Discrete Markov Bridge to "pull" theoretical designs into the synthesizable domain, identifying a sequence of reactions that approximate the target structure.
Steerable and Modular Synthesis
A key advantage of this direction is the ability to incorporate granular laboratory constraints. In the Saturn framework, Philippe Schwaller and Jeff Guo show how reinforcement learning can "steer" the generation process to avoid specific toxic reagents or to prioritize starting materials derived from industrial waste.
For high-speed discovery, models like ClickGen by Mingyang Wang and Shuai Li restrict the action space even further to "click chemistry"—highly reliable, modular reactions. This constraint doesn't necessarily sacrifice potency; the researchers demonstrated a 20-day turnaround from AI design to the wet-lab testing of nanomolar-level inhibitors, proving that baking synthesizability into the model drastically accelerates the transition from digital bits to physical atoms.
Why It Works
The intuition behind this research direction is that synthesizability is not a post-hoc filter, but a fundamental prior. By framing molecular discovery as a sequential decision-making task—where every "action" is a valid chemical transformation—the model is forced to stay within the bounds of physical reality.
As Sai Krishna Gottipati and Shengchao Liu noted in their work on PGFS, this doesn't just make the molecules makeable; it also makes the search more efficient. By reaching complex structures in fewer steps through multi-component reactions, these models reduce the variance of the search and focus computational power on the regions of chemical space that actually matter for drug development.